Porting my 1993 Amiga game to Godot, with an LLM reading the 68000 assembly
A 1993 Amiga game ported to the Godot engine using an LLM to read 68000 assembly has become a showcase for how modern models can reverse engineer and rebuild classic titles from binaries and ROMs. Commenters describe similarly successful ports across platforms like MSX, NES, DOS and PC, highlighting that models can extract assets, reconstruct game logic, and target modern frameworks with surprisingly little manual coding. Alongside the technical excitement, there’s debate over AI‑generated prose, authenticity, and whether such ports add value beyond emulation, but many see this as opening a new era of “software archaeology” and preservation for otherwise lost games.
LLM-assisted retro game porting
- Many commenters are excited that modern LLMs can read 68k assembly or old binaries and reconstruct games in engines like Godot or modern WebGL/WebGPU stacks.
- People report success with Amiga, MSX, ZX81, NES/SNES, Genesis, PC DOS titles, and more—sometimes getting 80–90% complete ports from a single long run, then iterating to fix assets and behavior.
- Several describe parallel efforts: generic “recompilation” frameworks, console-specific ecosystems (NES/SNES/Genesis/PlayStation/DS), and systematic pipelines for reverse-engineering and modernizing old games.
Workflows, tools, and limitations
- Claude Code is used to give the model direct access to directories containing binaries, old source trees, and target engine projects.
- Typical workflow: identify file formats, extract assets, unpack packed executables, disassemble (IDA or custom tools), segment code/data, then have the model reconstitute game logic in a modern language/engine.
- LLMs often need multiple passes and human steering; asset glitches and subtle behavior mismatches are common.
Debate over AI-authored prose
- Several readers find the blog’s “AI voice” off-putting and wish the writeup had been fully in the author’s own words.
- Others argue the story might not have been told at all without AI help, especially for a non‑native English speaker; editing AI drafts is seen as a pragmatic compromise.
- There’s concern that translation and assistance tools leave “LLM imprints,” making genuine posts look synthetic to readers and classifiers.
Original vs ported versions and authenticity
- Some argue there’s “no point” in porting if the original runs fine in emulators and prefer UAE/FPGAs for fidelity.
- Others counter that the port doesn’t subtract from the original, and AI assistance made a faithful, playable modern version feasible where a manual port would be too costly.
- The Amiga ADF release of the original game is noted as available for purists.
Technical fidelity and performance
- A small 108‑byte binary delta between old and newly assembled versions sparks discussion; plausible causes include assembler normalization and encoding differences, with behavior-level tests suggested as the real check.
- The Godot port deliberately avoids Godot’s physics to keep original gameplay feel; only minor tweaks (e.g., trampoline jump) were needed.
- Discussion covers PAL 50 Hz vs NTSC 60 Hz timing, Copper-based performance debugging, and curiosity about how a 512 KB assembly game compares to a modern high‑level engine in efficiency.
Nostalgia, history, and cost
- Commenters express awe at 1990s assembly‑only development under sanctions and thin documentation, and describe AI as “archaeology” for early personal computing.
- Multiple people share their own revived student or DOS projects, sometimes even ported to mobile or 3D.
- Cost is briefly discussed: one large “one shot” port reportedly used about half of a premium model’s weekly allowance; some users are consciously optimizing with cheaper models.